Dear adf-list,

going from adf-2000.02 to adf-2002 we consistently experience a 8-10 %
increase of CPU time needed for the same (geometry optimization) jobs. I
have been using the pre-compiled binaries on the very same AMD-athlon-750
machine in all cases. Notwithstanding changes in the actual code I suspect
the cause to be static linking with libraries that have not been optimized
for the specific cpu (686 or athlon). Unlike adf-2000, adf-2002 does run
on a classic pentium-133, which is a sign of high portability (but less
optimal code for more advanced cpus)

Hence, it may be worth the wile to recompile the source. Does anybody have
experiences with the run time behaviour of the adf-2002 code, which has
been recompiled with the most machine-specific options? We don't have a
source license, so we can't try ourselves

It should however be noted that we had some intermittent numerical
problems with the dynamically linked adf-2000 in the past, which could
only be solved with a statically linked executable (kindly supplied by
SCM). This is a price too high to be paid for speed.

regards

Hubert

+++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
Prof. Dr. Hubert Wadepohl
Anorganisch-Chemisches Institut der Universitaet
Im Neuenheimer Feld 270, D-69120 Heidelberg, Germany
e-mail bu9_at_ix.urz.uni-heidelberg.de
WWW http://www.uni-heidelberg.de/institute/fak12/AC/wadepohl
Phone (49) 6221-544827 FAX (49) 6221-544197
+++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
Received on 2002-04-18 16:55:35